Start here
Eight guided reading paths through the library. Pick the one that matches your math background and where you want to land. If nothing fits, the diagnostic calibrates a starting point from a 10-question placement.
Math to ML
From mathematical foundations through learning theory to why ML models work.
Assumes: Comfortable with calculus, linear algebra, and probability at the undergraduate level.
Open path →Engineer to ML Theory
Connect the tools you already use to the theorems behind them: when they work, when they break, and what the proofs actually say.
Assumes: You can train models and ship code; you want the math behind the methods.
Open path →Stats to Deep Learning
Bridge classical statistics to modern deep learning: from MLE and exponential families to transformers.
Assumes: Statistics background: MLE, hypothesis testing, regression, exponential families.
Open path →Master Linear Algebra
Linear maps, matrix operations, norms, eigenvectors, SVD, PCA, Jacobians, and matrix calculus. The algebra spine behind ML theory and neural networks.
Assumes: First-year calculus and basic matrix arithmetic. No prior linear algebra course required.
Open path →LLM From Scratch
A two-stage decoder-only path: next-token prediction and causal masking, then KV cache, FlashAttention, and modern inference.
Assumes: You can read PyTorch and have built a small neural network; you want to understand the GPT stack end to end.
Open path →Basic Neural Network From Scratch
Build a tiny MLP before jumping to transformers: linear layers, activations, losses, gradient descent, backprop, and regularization.
Assumes: Some Python and high-school calculus. No prior deep learning experience required.
Open path →ML Research Readiness
An advanced path through probability, learning theory, optimization, deep learning, and modern model behavior for stronger paper reading.
Assumes: Undergraduate probability, real analysis, and linear algebra; you want to read modern ML papers with fewer missing prerequisites.
Open path →Deep Learning Systems From Scratch
A 12-week shape, memory, and roofline track from linear layers to attention, KV cache, and accelerator performance.
Assumes: Comfortable with PyTorch, gradients, and backprop; you want the systems layer below the framework.
Open path →Other ways to navigate
- Reading paths — topic-spine sequences (concentration, generalization, frontier ML).
- Full curriculum — every topic in the graph, grouped by tier.
- Placement diagnostic — 10 questions to calibrate where to start.